overview-daodulisted

# overview-daodu ## Agent（claude / codex） 兩個 agent 皆可駕駛（Codex 端：`research-cards@private-plugins` plugin 的 overview-daodu skill）。讀卡一律優先用 `heptabase note read <id>`（或 obsidian 模式直接讀 vault 檔案）——不要依賴 mcp get_object（僅 Claude 有）。 ## Backend `heptabase`/`both`：原路徑不變。`obsidian`：`--card` 傳 vault 卡 id（`Folder/Name`）， 讀寫走 vault，導讀插入邏輯不變。 The unified `overview` skill's topics produce **structured but list-like** cards (paradigm tables, per-paper bullets, dimension tables). This skill adds the one thing those cards lack: a **teaching 導讀** at the very top that turns the pile of papers into a *story* — how the idea evolved, what problem each wave was answering — in the Hung-Yi Lee teaching voice. Division of labour: **you (the model) write the 導讀 prose**; the bundled `insert_daodu.py` does the mechanical insert/replace into the card. This is a **Mac-only** flow (local Heptabase + `heptabase` CLI). ## When To Use - User asks to add / 補 / 更新 / refresh a 導讀 / 導言 / 教學導讀 / teaching intro on an overview or comparison card. - After the `overview` skill adds new papers to a topic (missing 補洞 or an [elsewhere] facet fill) and the 導讀 should be re-synced to reflect them. ## Procedure 1. **Get the card id.** If the user named a topic (e.g. "Tokenizer/Codec (I)"), resolve it to the card id (the unified `overview` skill's `topics/<key>/config.py` exports the `OVERVIEW_CARD*` ids; the Key IDs table in `overview/topics/<key>.md` lists them too). Confirm with the user if ambiguous. 2. **Read the c